Pan-sharpening via Symmetric Multi-Scale Correction-Enhancement Transformers.

Area of Science:

• Remote Sensing
• Image Processing
• Computer Vision

Background:

• Pan-sharpening enhances remote sensing image quality for downstream tasks.
• Current deep learning methods often overlook image self-similarity, causing fusion artifacts and reduced clarity.
• Artifacts like ringing and poor spectral detail hinder the utility of fused high-resolution remote sensing images.

Purpose of the Study:

• To address limitations in existing pan-sharpening techniques by proposing a novel deep learning model.
• To improve the fusion of texture and spectral details in remote sensing images.
• To reduce artifacts and enhance the clarity of pan-sharpened images.

Main Methods:

• Introduction of the Symmetric Multi-Scale Correction-Enhancement Transformers (SMCET) model.
• Incorporation of a Self-Similarity Refinement Transformers (SSRT) module to capture intra-scale self-similarity in spatial and frequency domains.
• Utilization of an encoder-decoder framework for multi-scale transformations to simulate inter-scale self-similarity.

Main Results:

• SMCET demonstrates superior performance compared to existing pan-sharpening methods on multiple satellite datasets.
• The proposed model achieves enhanced fusion of texture and spectral details.
• Experimental results show a significant improvement in image clarity and a reduction in artifacts.

Conclusions:

• The SMCET model effectively addresses the limitations of existing deep learning pan-sharpening methods by incorporating self-similarity.
• The model's ability to capture self-similarity across multiple scales leads to superior image quality.
• SMCET offers a promising advancement for high-resolution remote sensing image enhancement.